Investigation and Fine Tuning of Hyper Parameters on U-Net Model for Segmentation of Glioma

  • Dheeraj D, Prasantha H.S
Keywords: Transfer Learning, Auto-Encoders, Glioma, Modified U-Net, Deep Learning, Segmentation.

Abstract

In medical Imaging, one of the most important aspects to be kept in mind is the preprocessing of the image before it is being deployed on different architectures for analysis. In recent past, it has been observed that in medical image diagnosis, deep learning models especially Convolution Networks have been declared as an efficient technique by most researchers. Since most of the times in biomedical imaging, the choice of the dataset and the region of interest plays a crucial role in diagnosis of the tumorous cell in human brain, an attempt has been made here to find the best variant of U-Net deep learning model for segmentation of the brain image as against the manual and automated approach of diagnosis of tumor in human brain. This paper investigates the different variants of the U-Net model before concluding the lighter and the best variant for the segmentation by taking into consideration fine-tuning of different hyper parameters involved in decision making. During the experimentation, the best model is selected not just, because one model outperformed the other in giving better accuracy, instead care has been ensured in selecting the model that has acceptable higher accuracy performance while performing fewer computations but giving out faster inference. In this work with the aid of transfer learning a block wise fine tuning of hyper parameters is carried out to derive a  model with 99.23% of accuracy on BraTs 2019 FLAIR dataset.
Published
2021-10-27
How to Cite
Prasantha H.S, D. D. (2021). Investigation and Fine Tuning of Hyper Parameters on U-Net Model for Segmentation of Glioma. Design Engineering, 7460-7473. Retrieved from http://thedesignengineering.com/index.php/DE/article/view/5787
Section
Articles